Bidirectional LSTM-CRF Models for Sequence Tagging

The paper systematically proposes and compares a family of Long Short-Term Memory models for sequence tagging: plain LSTM networks, bidirectional LSTM (BI-LSTM), LSTM with a Conditional Random Field layer (LSTM-CRF), and bidirectional LSTM with a CRF layer (BI-LSTM-CRF). This is the first work to apply the BI-LSTM-CRF model to standard NLP sequence tagging datasets, combining a bidirectional LSTM that uses both past and future input features with a CRF layer that exploits sentence-level tag information.

The BI-LSTM-CRF model produces state-of-the-art or near state-of-the-art accuracy on part-of-speech tagging, chunking, and named entity recognition benchmarks. In addition, it proves robust and shows less dependence on word embeddings than previously observed, making it a strong and practical architecture for neural sequence labeling.